arena

并行候选，选 base，再嫁接

Fan out N parallel attempts at the same task. Read every candidate end to end. Pick the strongest as the base. Graft the best ideas from the others into it. Verify the synthesized result.

对同一任务扇出 N 次并行尝试。通读每个候选。选最强的做 base。把别人最好的想法嫁接进去。验证综合结果。

Start

开始

Open a todolist with one entry per phase before launching anything.

启动前打开 todolist，每个阶段一条。

1. Frame

2. Fan out

3. Cross-judge

4. Pick

5. Graft

6. Verify

7. Frame（定框）

8. Fan out（扇出）

9. Cross-judge（交叉评判）

10. Pick（选 base）

11. Graft（嫁接）

12. Verify（验证）

Phase A: Frame

Phase A: 定框

The N candidates will receive the same prompt, so the prompt is the contract.

N 个候选拿到同一份 prompt，所以 prompt 就是契约。

1. State the artifact each candidate is producing.

2. Derive the rubric. State what success looks like for this task, then turn it into 3-6 concrete gradeable criteria. The rubric is the picker’s tool in Phase D. Candidates only see the task.

3. Pick the runners. Use arena runners from ~/.cursor/rules/pstack-models.mdc when present. Otherwise default to one each on claude-opus-5-5-max, gpt-5.6-sol-max, grok-4.7-xhigh-fast. Spawn more when the arena covers multiple design directions. Same model N times when the work is generation-bound rather than judgment-sensitive.

4. Assign output paths. Each candidate writes to its own location (a git worktree where possible, otherwise /tmp/arena-<slug>/candidate-<n>/), per the separate-before-serializing-shared-state principle skill.

5. 说清每个候选要产出什么 artifact。

6. 推导 rubric。先说这次任务怎样算成功，再变成 3–6 条可打分的具体标准。rubric 是 Phase D 挑选者的工具；候选只看到任务本身。

7. 选 runners。有 ~/.cursor/rules/pstack-models.mdc 里的 arena runners 就用；否则默认各一个：claude-opus-5-5-max、gpt-5.6-sol-max、grok-4.7-xhigh-fast。arena 覆盖多个设计方向就多 spawn。工作偏生成而非判断敏感时，同一模型跑 N 次。

8. 分配输出路径。每个候选写自己的位置（能用 git worktree 就用，否则 /tmp/arena-<slug>/candidate-<n>/），按 separate-before-serializing-shared-state principle skill。

Phase B: Fan out

Phase B: 扇出

Spawn all N subagents in one message with run_in_background: true, each with the task, the path to the shared grounding, its own output path, and instructions to produce both the artifact and a short rationale.

在一条消息里 spawn 全部 N 个 subagent，run_in_background: true；各自带任务、共享摸底材料路径、自己的输出路径，以及「产出 artifact + 短 rationale」的说明。

Each rationale names the alternatives the candidate considered and what it rejected.

每份 rationale 点名候选考虑过的备选，以及它拒绝了什么。

If a candidate fails to produce output, proceed with N-1 and note the dropout in the synthesis record.

某个候选没产出，就用 N-1 继续，并在综合记录里记下脱落。

Phase C: Cross-judge

Phase C: 交叉评判

After all Phase B candidates complete, choose one model from the arena cross-judge pool in ~/.cursor/rules/pstack-models.mdc when present. Otherwise use claude-opus-5-5-max, gpt-5.6-sol-max, grok-4.7-xhigh-fast. Prefer a different model family from the parent’s. Spawn one readonly judge subagent on that model. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with the parent’s reading in Phase D, not with the candidates themselves. Don’t spawn the judge while candidates are still writing.

Phase B 全部完成后，从 ~/.cursor/rules/pstack-models.mdc 的 arena cross-judge pool 选一个模型（没有就用 claude-opus-5-5-max / gpt-5.6-sol-max / grok-4.7-xhigh-fast）。优先与 parent 不同模型族。在该模型上 spawn 一个只读 judge subagent。它看到 rubric 和按路径标签的候选，逐条打分，并推荐 base + rationale。它与 parent 在 Phase D 的阅读并行，不与候选本身并行。候选还在写时别 spawn judge。

Phase D: Pick a base

Phase D: 选 base

Read every candidate end to end before picking.

挑选前通读每个候选。

Score each candidate against the rubric criterion by criterion, not on holistic feel. Compare against the cross-judge. Agreement on the base confirms the pick. Disagreement means one of you is biased or the rubric was ambiguous. Read both rationales before deciding.

按 rubric 逐条打分，别凭整体感觉。对照 cross-judge。base 一致就确认；不一致说明有人偏了或 rubric 含糊。决定前读双方 rationale。

Pick the base on which candidate a future maintainer can extend most easily without breaking invariants. Prefer the cleaner boundary or smaller API when two feel tied, per the Laziness Protocol.

选那个未来维护者最容易扩展、又不破坏不变量的候选做 base。打平时优先更干净的边界或更小的 API，按 Laziness Protocol。

Record the pick and the reason in a short synthesis note alongside the base artifact, including the cross-judge’s verdict.

在 base artifact 旁写短综合笔记：选了谁、为什么，含 cross-judge 的裁决。

Phase E: Graft

Phase E: 嫁接

Walk each losing candidate once more and identify what is worth porting into the base. The signal is usually one or two things per candidate, not most of it.

再扫一遍落选候选，找出值得迁入 base 的东西。信号通常是每个候选一两处，不是大半。

Fold each graft in by hand, per the redesign-from-first-principles principle skill. Don’t paste mechanically. The result has to remain coherent under one mental model.

按 redesign-from-first-principles 手工折入每处嫁接。别机械粘贴。结果必须在同一心智模型下连贯。

Record what was grafted, from which candidate, and what was rejected and why.

记录嫁接了什么、来自哪个候选、拒绝了什么及原因。

When N candidates converge on the same shape, that is a strong agreement signal. Note the convergence in the record and ship the consensus shape. No graft is needed. When N candidates wildly diverge, Phase A was under-specified. Reframe and re-run rather than averaging the divergence.

N 个候选收敛到同一形状，是强一致信号。记入记录，交付共识形状，不必嫁接。若严重发散，是 Phase A 定框不够。重新定框再跑，别把分歧平均掉。

Phase F: Verify

Phase F: 验证

The synthesized artifact has to hold up under the same scrutiny as any other output, per the prove-it-works principle skill.

综合产物要经得起与其他产出同等的审视，按 prove-it-works principle skill。

If verification surfaces a problem the arena did not catch, either Phase A was wrong (re-frame and re-run) or one candidate caught it and you missed the graft (go back to Phase E). Don’t paper over.

若验证露出 arena 没抓到的问题：要么 Phase A 错了（重定框再跑），要么某个候选抓到了你漏嫁接（回 Phase E）。别糊弄过去。

Outputs

产出

One synthesized artifact. One short synthesis note alongside, naming the base, the grafts (with source candidate), the rejections, the dropouts if any, and the verification result.

一份综合 artifact。旁附短综合笔记：base、嫁接（带来源候选）、拒绝项、若有脱落、以及验证结果。